4.6 AI Tools (Summary, Mark, PII Blur)

4.6.1 The Three Tools, in One Sentence Each

All three respect the non-negotiable privacy rule: the server never receives unencrypted message bodies or PII.

4.6.2 Chat Summary

What the user sees

A button in the Chatraum Einstellungen menu:

🪶 Chat Zusammenfassen — "Spare Zeit, mit Hilfe unseres vollends Datenschutzkonformen KI Workflows."

Click → side panel with a generated summary, editable, savable as a counselor-owned case note.

How it actually works

sequenceDiagram
  autonumber
  participant Co as Counselor
  participant FE as ORISO-Element
  participant AI as Self-hosted AI Workflow
  participant US as UserService
  Co->>FE: Click Chat Zusammenfassen (⇧Ü)
  FE->>FE: Decrypt last N events locally (Megolm)
  FE->>FE: PII-scrub (names, emails, account numbers, phone)
  FE->>AI: POST /summarize { scrubbed_text, language }
  AI-->>FE: Bullet-point summary
  FE->>Co: Show in editable side panel
  Co->>FE: Edit, Save
  FE->>US: PUT /case-notes/{id} { encrypted_at_rest, owner=counselor }
  US-->>FE: 200 OK (opaque session_id link only)

Hard rules

• Decryption happens only in the browser.
• The AI service is run inside the cluster (self-hosted). No third-party LLM calls in production.
• Output is counselor-owned, not client-owned — survives client wipe.
• The note saves only an opaque session_id link, never a pseudonym after wipe.
• Counselor can disable AI for any room; admin can disable it tenant-wide.

Edge cases

• AI service down → button greys out, counselor falls back to manual note. No crash, no data loss.
• Window of unread events → counselor picks how many last-N events to summarise; default = whole open session.
• Mid-session run → Summary marked as "interim" so the counselor can refresh later.

4.6.3 Mark Text

What it is

A Tiptap-style highlighter on every message. The counselor selects text, picks a color from a palette, and the marked range is preserved as an annotation — independent of the original encrypted body.

Use cases

• Qualitative coding: tag sentences as "Great Aspects" (green), "Dangerous Behaviors" (red), "Step 1 / 2 / 3 …", etc.
• Build a case note: send all marked text to the side editor.
• Build a thread: send a single marked excerpt to a side-thread for team discussion.

UX vocabulary (from Figma)

How it actually works (proposed)

Marks are annotation events in Matrix, separate from the message event:

{
  type: "de.oriso.annotation.mark",
  content: {
    target_event_id: "$abc:matrix.oriso-dev.site",
    range: { start: 12, end: 47 },
    color: "#FFD54F",
    label: "Great Aspects",
    blur: false,
    visibility: "private_to_author"  // or "shared_in_thread"
  }
}

The annotation is itself encrypted (Megolm). Receivers fetch annotations alongside messages and overlay them client-side. This keeps marks portable across re-keys and avoids polluting the original message.

Status

[NEW FEATURE] — needs the annotation event type to be added to the Element fork (ORISO-Element).

4.6.4 PII Blur

The same flow as Mark Text but with a blur flag set on the range.

Visual specification

• Uniform white overlay at 80% alpha, blur radius 4 px (per Figma label fff 80% Overlay Blur (uniform) 4).
• Overlay covers the exact selected range; receiver sees an opaque rectangle.
• A system message accompanies the first blur: "Counsellor blurred some personal please be aware in future."
• The eye toggle in the counselor's own view lets them peek under the blur (since they own the original); receivers cannot.

Who can blur

• Counselor: yes (default).
• Supervisor: yes (so they can teach by example).
• Client: no (clients should not be redacting their own counselor's text).

How it actually works

The blur metadata is part of the same annotation event as Mark Text, with blur: true. Receivers' clients honour the flag at render time. Bodies remain encrypted; the blur is purely a presentation rule.

Edge cases

• Blur on a message that gets deleted → annotation is orphaned; client garbage-collects.
• Blur on a thread that gets archived → annotation persists with the chat; deleted with the chat.
• Receiver has older client without blur support → fallback: render the body in plain text but show a banner "This message contains content the counselor wished to blur — please disregard." (Decision needed — see Figma analysis §4.4.)

4.6.5 Per-User Visibility — "Message Visible to..."

Both Mark and Blur can be combined with the Multi-Recipient Send (see Group Chats §4.5.4) so that a marked, blurred excerpt is visible only to the moderators while the original message remains visible to all.

4.6.6 Permissions

These tools are counselor-only by default and can be globally toggled by admins. The Permissions Matrix (see Roles & Permissions) controls them per chat-type.

Admins (Tenant / Agency) can disable any of these per chat type to satisfy strict-privacy customers.

4.6.7 Hard Rules (Reaffirmed)

• ❌ Server never sees plain text.
• ❌ AI service is never third-party in production.
• ❌ Marks / Blurs never leak the underlying body to non-counselor receivers.
• ✅ Counselor can edit any AI-generated content before saving.
• ✅ Save retains only opaque session_id link after client wipe.
• ✅ Tenant / Agency admin can disable any tool with one click.

4.6.8 Related

• Group Chats (4.5) — multi-recipient send pairs with these tools.
• Roles & Permissions — per-chat-type matrix.
• Edge Cases — failure modes for AI, blur, marks.
• Figma Analysis — source screens.